Maximum Power Point Tracking for Photovoltaic Optimization Using

IEEE Transactions on Power Electronics:
Accepted 25-Apr-2010
Maximum Power Point Tracking for Photovoltaic Optimization
1
Using Ripple-Based Extremum Seeking Control
Steven L. Brunton1 , Clarence W. Rowley1 , Sanjeev R. Kulkarni1 , and Charles Clarkson2
1 Princeton
2 ITT
University, Princeton, NJ 08544
Geospatial Systems, Rochester, NY 14606
Corresponding Author: Steven Brunton
e-Mail: sbrunton@Princeton.Edu
Phone: (609)-921-6415
Fax: (609)-258-6109
Address: Steven Brunton
MAE Department
D-wing, Engineering Quadrangle
Princeton University
Princeton, NJ 08544
Parts of this work were presented at the 34th IEEE Photovoltaic Specialists Conference in Philadelphia on June 7-12, 2009.
Keywords: Maximum power point tracking, extremum seeking, rapidly varying irradiance, inverter ripple, photovoltaic power
generation
Abstract
This work develops a maximum power point tracking (MPPT) algorithm that optimizes solar array
performance and adapts to rapidly varying irradiance conditions. In particular, a novel extremum seeking
(ES) controller which utilizes the natural inverter ripple is designed and tested on a simulated solar array
with a grid-tied inverter. The new algorithm is benchmarked against the perturb and observe (PO) method
using high-variance irradiance data gathered on a rooftop array experiment in Princeton, NJ. The extremum
seeking controller achieves efficiencies exceeding 99% with transient rise-time to the MPP of less than .1
seconds. It is shown that voltage-control is more stable than current-control and allows for accurate tracking
of faster irradiance transients. The limitations of current-control are demonstrated on an example. Finally,
the effect of capacitor size on the performance of ripple-based extremum seeking control is investigated.
I. INTRODUCTION
2
Recently, there has been significant environmental and political motivation to shift domestic power generation to renewable sources such as wind and solar. Solar power is at the forefront of clean, renewable energy, and it is gaining momentum
due to advances in solar panel manufacturing and efficiency as well as increasingly volatile fuel costs. Solar power is an
attractive option because of the large amount of power available in incident sunlight, particularly in large industrial parks
and residential suburbs. However, photovoltaic (PV) solar cells, the most readily available solar technology, operate best
on bright days with little or no obstruction to incident sunlight. Frequent overcast days and partial obstructions such as tree
limbs or neighboring buildings limit the efficiency and reliability of solar power throughout much of the United States [1].
This work is motivated by the need to optimize solar array performance for rapidly varying environmental conditions, such
as those characteristic of New Jersey’s climate.
Because of the photovoltaic nature of solar panels, the current-voltage, or I -V , curves depend nonlinearly on temperature
and irradiance levels [2], [3]. Therefore, the operating current and voltage which maximize power output will change with
environmental conditions, as in Fig. 2. In order to maintain efficient operation despite environmental variations, one approach
is to use a maximum power point tracking (MPPT) algorithm to dynamically tune either control current or voltage to the
maximum power operating point.
Typically MPPT algorithms are implemented on a solar array using a switching power converter. For example, with a gridtied inverter, the solar array charges a capacitor, and then current is switched out of the capacitor at a varying duty cycle in
order to reconstruct a sinusoidal current which injects power into the grid. A number of solar power converter architectures
are discussed in the literature [4], [5], [6], [7], [8]. In the majority of power converters, the internal switching mechanism
imposes a voltage and current ripple which is felt by the PV array. Minimizing the magnitude of this ripple has been a major
concern, and is achieved by careful (and expensive) choice of the capacitor and inductor.
There are a number of maximum power point tracking algorithms for changing environmental conditions [9], [10]. Control
algorithms which do not assume a particular model and are adaptive to changing system parameters are ideal for a number
of reasons, including less frequent maintenance and fine tuning. Adaptive, model independent algorithms are applicable to
a wide range of panel and inverter technologies. A number of “black box” MPPT algorithms, such as perturb and observe [9]
and incremental conductance [2] have been explored in the literature [10].
Perturb and observe (PO) is a workhorse MPPT algorithm because of its balance between performance and simplicity. In
its simplest form, the perturb and observe algorithm tracks the MPP by perturbing the control input in a given direction and
observing if the output power goes up or down; if the power increases, the perturbation direction is unchanged, and if the
power decreases, the direction is reversed. Because the standard perturb and observe method uses a fixed perturbation
3
size, it suffers from a performance tradeoff between transient rise-time and steady-state performance. PO has also been
shown to track in the wrong direction given rapidly varying irradiance [10]. Modified versions of PO have been proposed,
for example by using an adaptive step which is related to ∆P/∆u (where P is power and u is either perturbed voltage or
current) [11] or by including rudimentary model assumptions [12].
A promising new MPPT algorithm is the method of extremum seeking (ES) control, which may be closely related to the
ripple correlation control (RCC) and PO methods. The extremum seeking method of Krstic [13] offers fast convergence
and good steady-state performance with guaranteed stability for a range of parameters. [14] and [15] implement extremum
seeking control by injecting an external perturbation signal. Ripple correlation control (RCC) utilizes the natural inverter
ripple and corrects the duty cycle of the switching converter in order to set ṖI˙ = 0, which is a condition for the MPP. The
duty cycle is updated using either a discrete comparison, as in perturb and observe [21], [22], [16], or using the product of
the high-pass filtered power and control variable, as in extremum seeking control [17], [18].
The approach here is to develop an extremum seeking controller that utilizes the natural 120Hz inverter ripple to track the
MPP in rapidly varying irradiance conditions. In particular, the extremum seeking controller is compared against a well-tuned
PO algorithm on high-variance irradiance data measured for use on a solar array on the roof of Princeton’s engineering quad.
This work also compares the use of voltage and current as control variables and demonstrates limitations of current-control
for gradient-climb methods. Finally, the effect of capacitor size on extremum seeking control is investigated.
II. PV ARRAY INVERTER MODEL
In order to simulate a comparison of various MPPT algorithms, it is necessary to model the array-inverter dynamics.
Figure 1 is a schematic of the array-inverter system. The block labelled PV Array encapsulates all of the dynamics associated
with the solar array, including the functional dependence of the I -V curves with irradiance G and temperature T . The
switching dynamics of the inverter are encapsulated in the block labelled u, and are discussed below. L and C represent the
inductor and capacitor, respectively.
VL
uI
uV
IC
VC
!"# I
$%%&' V
Fig. 1. Schematic of PV array and inverter with LC dynamics. The inverter control variable u is either current or voltage and is
controlled by the inverter’s switching logic.
4
A. PV Array Model
The array I -V curve may be written I = I(V, G, T ) and is modeled by the lighted-diode equations [2], [3]:
I
IOS
IL
�
= IL − IOS exp
= IOR
�
T
TR
�
�
q
(V + IR) − 1
AkB T
�
qEG
exp
AkB
�
1
1
−
TR R
(1)
��
(2)
G
(ISC + KT,I (T − TR ))
1000
=
(3)
where I and V are the same as in Figure 1, IL is the light generated current, IOS is the cell reverse saturation current, and T
is temperature. Because temperature variations are typically much more gradual than irradiance changes, we assume that
temperature is constant, Tconst = 300K , for the remainder of the analysis. Thus the functional dependence on T is dropped.
V = V (I, G) � V (I, G, Tconst )
(4)
I = I(V, G) � I(V, G, Tconst )
(5)
are equivalent representations of the I -V curves for a constant temperature Tconst and changing irradiance G. Values and
definitions of other terms in the equations are given in the Appendix.
Figure 2 shows the I -V , P-V and P-I curves for varying irradiance G using the above equations and parameters. The P-V
characteristic curves are more symmetrical about the maximum power input than the P-I curves. Moreover, the maximum
power point occurs at a smaller normalized voltage input, giving voltage-control larger failure margins.
I−V Curves for Varying Irradiance G
P−V and P−I Curves for Varying Irradiance G
15
5000
P−V curve
P−I curve
4000
Power (P)
Current (I)
10
3000
2000
5
1000
0
0
50
100
150
200 250
Voltage (V)
300
350
400
0
0
0.2
0.4
0.6
Normalized Input (V or I)
0.8
1
Fig. 2. (left) I-V curves at several irradiance levels from G = 50 W /m2 to G = 1000 W /m2 (spaced 50 W /m2 apart). (right) P-V
(black) and P-I (gray) curves plotted against an input variable which is normalized by its maximum allowable value. The maximum
power point is labeled as either a black circle or gray diamond.
5
B. Inverter Model
Applying Kirchoff’s law to the circuit in Figure 1 yields the following relationships:
I
uV � VC
= uI + IC
(6)
= −V −VL
(7)
where the inverter control variable u is either current, uI , or voltage, uV .
If the inverter control variable is current, uI , the array I -V curve has the form V = V (I, G) and equation (7) becomes:
VC
=⇒
−
dVC
dt
dI
dt
d
d2I
V (I, G) + L 2
dt
dt
= −V (I, G) − L
(8)
=
(9)
=
∂V (I, G) dI ∂V (V, G) dG
d2I
+
+L 2
∂I
dt
∂G
dt
dt
(10)
Equation (6) and the capacitor equation yield:
dVC
IC
=
dt
C
=⇒
dVC
1
= − (uI − I)
dt
C
(11)
Combining equations (10) & (11) yields the system dynamics in terms of inverter control current uI and array current I :
LC
d2I
∂V dI
∂V dG
+C
+ I = uI −C
2
dt
∂I dt
∂G dt
(12)
The dynamical system given by equation (12) represents a forced oscillator with nonlinear damping. The forcing corresponds to the inverter control current uI as well as the change in I -V curve due to irradiance change, given by −C(∂V /∂G)(dG/dt).
To flow 60Hz AC power into the grid at a given current ûI , the inverter switches DC current out of a large capacitor. This
requires the following inverter control current with a large 120Hz oscillation:
uI = ûI (1 + sin(120 × 2πt))
(13)
In practice, the LC circuit acts as a low pass filter between the control current uI , and the array current I , so that I
experiences a 120Hz ripple at approximately 3% magnitude:
I ≈ ûI (1 + .03 sin(120 × 2πt + ϕ))
6
(14)
There is also a high frequency ripple at 20kHz due to the inverter sampling time; however, this has a negligible effect and
is not modeled. For more information on DC-AC power inverters, see Bose [5].
Similarly, if the inverter control variable is voltage, uV , the array I -V curve has the form I = I(V, G), which yields the
following voltage-control dynamics
L
∂I
∂I
V̇ +V = −uV − L Ġ
∂V
∂G
(15)
Notice that the system type and order of the current-control dynamics (12) and voltage-control dynamics (15) are different.
This difference is due to the fact that the control voltage is instantaneously reflected in the capacitor voltage for systems of
the form given in Figure 1.
III. MAXIMUM POWER POINT TRACKING
In the maximum power point tracking (MPPT) algorithms implemented below, it is assumed that there are current and
voltage measurements which may be multiplied to obtain a power measurement. The control variable is either current or
voltage, which are determined by the MPPT algorithm and commanded by the inverter. The algorithms do not require any
additional sensors or models to track the MPPT despite exogenous disturbances due to changing irradiance, G(t).
Although there are a number of MPPT algorithms that have been developed over the past two decades, this analysis
compares a new method, ripple-based extremum seeking (ES) control, with the standard method, perturb and observe
(PO). Additionally, an interesting relationship between the existing ripple correlation control (RCC) method and extremum
seeking control is demonstrated.
The efficiency of each method is given by the formula
R
t
Pactual (τ)dτ
ηMPPT (t) = R0 t
0 Pmax (τ)dτ
(16)
and the transient rise-time, τrise , is the time it takes the algorithm to reach 90% of the instantaneous maximum power point.
A. Perturb and Observe
The most prevalent MPPT algorithm is perturb and observe (PO). PO repeatedly perturbs the input by a fixed amount
in a given direction, and the direction is changed only if a drop in power is detected. Although this algorithm benefits from
simplicity, it lacks the speed and adaptability necessary for tracking fast transients in irradiance.
7
A simple variant of the PO method uses an average power for the comparison step that has been averaged over a
number of inverter switching cycles. The effect of this averaged power comparison is a decrease in noise, and also that the
perturbation magnitude no longer must exceed that of the natural ripple. This is the version of PO used in this comparison.
Finally, it is possible to construct a PO algorithm with an adaptive step
∆uk+1 = ∆uk +
∆Pk
∆uk
(17)
This overcomes the tradeoff between transient rise and steady-state tracking efficiency. However, there are no guaranteed
stability conditions for this algorithm, and it is not implemented in this study.
B. Extremum Seeking Control
A new adaptive maximum power point tracking algorithm is based on the extremum seeking (ES) control method. A
schematic of the algorithm is shown in Figure 3. This controller converges at a rate which is proportional to the slope
of the power curve, either P-I or P-V , and has guaranteed stability over a range of system parameters [13], [20]. The
algorithm works by adding a perturbation signal α sin ωt to the “best guess” for the input û which maximizes the quantity
of interest, namely the array output power. The perturbation passes through the system and produces a perturbation in the
power. Multiplying the high-pass filtered control variable uhp (u is uI or uV ) and high-pass filtered power Php results in a
demodulated signal ξ which is positive to the left of the MPP and negative to the right of the MPP. Integrating this signal and
adding it to the best guess û causes the control variable û to adaptively track the MPP. A simple averaging analysis, which
assumes constant environmental conditions, shows that the demodulated signal ξ is proportional to the slope of the power
curve:
Z
ω 2π/ω
uhp · Php dτ
2π 0
∂P
≈
(û, G) · α2 sin2 ωt.
∂û
ξavg =
(18)
(19)
In practice, rather than injecting a sinusoidal control perturbation α sin ωt , as in the standard extremum seeking algorithm,
it is convenient to utilize the inverter ripple for the perturbation. Using the control signal in equation (13), the array current and
power will have a small ripple, as in equation (14). Therefore, the high-pass filtered array current and power are multiplied,
yielding the demodulated signal, ξ, similar to Figure 3. A schematic of the new ripple-based extremum seeking algorithm is
shown in Figure 4. Details regarding how to tune the extremum seeking controller are given in the Appendix.
2"33$+0&
24%%-+.
Iˆ
+
"4.54.&
5"1-%
!"#$%&'%%$(&)
*+,-%.-%&/$%01$%-
ξ
k
s
×
ρ
8
s
s+ωh
7+.-8%$."%
678695$::&
!#.-%
α sin ωt
Fig. 3.
Extremum seeking algorithm.
%+44!(2$
%&""'()
×
1 + sin ωt
Fig. 4.
û
*+,!"$-""!#$.
/(0'")'"$1!"23!"'
ξ
k
s
6()'7"!)+"
Ihp
×
!""!#$5+3'"
!""!#$%&""'()
s
s+ωh
867895!::$
!,)'":
Php
s
s+ωh
Extremum seeking algorithm utilizing natural inverter ripple.
C. Ripple Correlation Control
Ripple correlation control (RCC) utilizes the natural inverter ripple to either perform perturb and observe [21], [22], [16]
or extremum seeking control [17], [18]; RCC has recently been extended to a discrete-time formulation [19]. RCC corrects
the duty cycle of the switching converter according to the integral:
d(t) = k
Z
dP dI
· dt ≈ k
dt dt
Z
Php · Ihp dt
(20)
where Php and Ihp are first-order high-pass filtered quantities that approximate the derivatives in Eq. (20), and k is the
integrator gain, similar to Figures 3 and 4. Using this varying duty cycle, RCC tracks the maximum power point by tracking
the equivalent condition ṖI˙ = 0. In practice, this implementation of RCC is similar to the ES algorithm with first order highpass filters. However, RCC lacks the mathematical foundation and careful stability analysis that makes extremum seeking
attractive for control design. RCC is limited to first order high-pass filters, which approximate the derivative at low frequencies,
whereas ES also works with higher order filters.
IV. SIMULATED MPPT COMPARISON
This section summarizes the results of a simulated maximum power point tracking (MPPT) comparison between the
extremum seeking (ES) and perturb and observe (PO) algorithms. The algorithms are tested on rapidly varying irradiance
9
data from measurements on an partially cloudy day. Temperature is assumed constant throughout the simulations because
variations in temperature are more gradual than in irradiance. This simplifies the dynamic models (12) and (15). Two sets of
algorithm comparisons are simulated, depending on whether voltage or current is controlled by the inverter.
Using the irradiance measurements as an input to the model, both the extremum seeking (ES) and perturb and observe
(PO) algorithms are tested. The extremum seeking model is shown in Figure 4. Information about the simulated PV system
are given in the Appendix.
A. Irradiance Data
Figure 5 shows irradiance data for two consecutive days in June, 2007. The data was measured on a rooftop of the
Engineering Quadrangle at Princeton University. Irradiance is more erratic on the second day because of rapidly moving,
scattered cloud cover. The bottom of Fig. 5 provides a detailed view of the irradiance data over a 25 minute window between
12:34-12:59AM on June 20th , 2007 (day 2). This time range is chosen because it includes rapid irradiance changes, and
because a short 25 minute window makes it possible to see the controller response to individual irradiance changes.
Irradiance (W/m2)
1400
day 1
day 2
1200
1000
800
600
400
200
0
5
10
15
20
1000
1500
Irradiance (W/m2)
Time (hr)
1200
1000
unfiltered
lowïpass
800
600
0
500
Time (s)
Fig. 5. (top) Irradiance data for two days in June in Princeton, NJ. (bottom) Irradiance data spanning 25 minutes from 12:3412:59AM on June 20th , 2007. Signal is low-pass filtered (black), so that noisy measurements (gray) are averaged over about 10
seconds.
The noisy irradiance measurements are low-pass filtered so that the data is averaged over about 10 seconds. Because
10
the irradiance sensors are located at points along the array, the measured irradiance responds more quickly to disturbances
than the array. Determining the actual time scales on which irradiance changes affect a solar array will depend on the size
and orientation of the array, and should be investigated further. Both unfiltered and low-pass filtered irradiance sets are used
in the simulations below.
B. Voltage-Control
Figure 6 shows the results from a MPPT comparison where the inverter control variable is voltage. The array power,
current and voltage are plotted in time for the ES (black) and PO (light gray) algorithms as well as the true maximum power
(dark gray). Both PO and ES command a control voltage which oscillates closely around the true maximum power voltage,
as seen in the bottom plot. The extremum seeking method accurately tracks the maximum power point and rises to the MPP
orders of magnitude more rapidly than perturb and observe.
In the voltage-control simulation, each algorithm tracks the MPPT of the unfiltered irradiance measurements. The unfiltered irradiance data varies more rapidly than the filtered data, and therefore the MPPT is more difficult to track. However,
voltage-control results in almost perfect MPPT with both ES and PO achieving efficiencies around 99%. ES has efficiency
ηES = .9968 and rises to the MPP in .1s, and PO has efficiency ηPO = .9939 and rises to the MPP in 6s. The inverter
hardware uses a 2000µF capacitor and 1µH inductor. The PO step size is ∆V = 5V with 1kHz sampling. Finally, ES utilizes
the natural inverter ripple, which has magnitude 3%.
C. Current-Control
Figure 7 shows the results from a MPPT comparison where the inverter control variable is current. The array power,
current and voltage are plotted in time for the ES (black) and PO (light gray) algorithms as well as the true maximum power
(dark gray). The extremum seeking method commands a control current which oscillates closely around the true maximum
power current, as seen in the middle plot.
Capacitor Size (µF)
ES Efficiency (%)
250
90.71
500
96.79
1000
98.99
2000
99.63
3000
N/A
TABLE I
E XTREMUM S EEKING EFFICIENCY FOR VARYING INVERTER CAPACITOR SIZES .
In the current-control simulation, each algorithm tracks the MPPT of the low-pass filtered irradiance data. The currentcontrol dynamics are not easily controllable and fail for the rapid irradiance changes found in the unfiltered data. However,
on the low-pass filtered data, current-control with PO and ES admits almost perfect MPPT. ES has efficiency ηES = .9963
with rise-time of .02s, and PO has efficiency .9898 with rise-time of 6.9 seconds. The step size for PO is ∆I = .5A.
11
Although both ES and PO track the maximum power point with high efficiency, ES is a more adaptive algorithm. This
is seen in Fig. 8, where a number of PO algorithms are compared on an array-inverter system with a 2000µF capacitor.
The incremental step is tested using the values {.05A, .1A, .25A, .5A, 1.0A}. The perturb and observe step size ∆I = .5A
is chosen because this is the only value that doesn’t cause the algorithm to fail. For small ∆I the PO algorithm suffers from
slow rise-time and insufficient speed to track large irradiance changes, resulting in failure. For the largest step ∆I = 1.0A,
the rise-time is faster, but the step size is so large that the oscillation about the MPP reaches the short-circuit current.
In contrast, the ES controller performs well for a large range of ripple magnitudes. To compare the ES algorithm for a
number of different ripple magnitudes, it is necessary to vary capacitor size from 250µF to 3000µF ; larger capacitors result
in smaller ripple, and vice-versa. The comparison is shown in Fig. 9. As the capacitor size increases, the ES algorithm tracks
the MPP with less oscillation due to the decrease in ripple magnitude. However, if the capacitor is too large the ripple-based
control signal is insufficient to track irradiance changes, and the algorithm fails as seen in the case with a 3000µF capacitor.
However, as seen in Tab. I, ES performs at high efficiencies even for small capacitor systems with large ripple.
D. Limitations of Current-Control
As irradiance decreases rapidly, the I -V curve shrinks and the maximum power voltage (MPV) and current (MPI)
decrease. If the MPPT algorithm does not track fast enough, the control current or voltage will “fall off” the I -V curve.
Figure 10 shows irradiance plotted against the MPV and MPI as well as the open-circuit voltage and short-circuit current.
From these plots, it is clear that voltage-control will benefit from much larger margins given rapidly decreasing irradiance.
Moreover, if the MPPT algorithm “falls off” the I -V curve, in the case of voltage-control this corresponds to open-circuit, and
in the case of current-control this corresponds to short-circuit. Therefore, voltage-control not only provides safer margins of
operation, but the failure mode is more acceptable than in the case of current-control.
Figure 12 shows the current-control extremum seeking algorithm failing to track a rapid irradiance change when the ripple
is small because of a large capacitor, C = 3000µF. The figure on the left shows the power tracking in time, and the figure
on the right shows the corresponding P-I tracking. After a fast rise (1), the algorithm tracks the MPP with increasing (2) and
decreasing (3) irradiance until the algorithm fails (4). Between steps (3) and (4) the control current does not track the MPP
quickly enough, and finally the control current falls off of the I -V curve, resulting in a short-circuit.
Finally, Fig. 11 shows the high-pass filtered current Ihp and the demodulated product ξ. The signal ξ is integrated into
the algorithm’s “best guess” of where the MPP current is. At the point of failure, the magnitude of the signal ξ goes to zero.
This is explained by the high-pass filtered current Ihp which also goes to zero at the point of failure. Because the current is
tracking slowly, as irradiance falls the current gets closer to the short-circuit current, and therefore the ripple is constrained
by the hard wired short-circuit current. For this reason, as the current gets closer to Isc , the magnitude of the current ripple
in the array is constrained and goes to zero, causing the controller to fail.
12
E. Step Irradiance Changes
It is useful to consider the dynamic response of each method to step changes in irradiance. In addition to representing
a worst-case scenario, step irradiance changes provide a controlled disturbance on which to compare the tracking performance of each method. Figure 13 shows the response of each method to step changes in irradiance. The irradiance starts
at 200W /m2 at time t = 0s and steps up to 1000W /m2 at t = 20s. At t = 40s the irradiance steps back down to 200W /m2 .
Extremum seeking (ES) and perturb and observe (PO) are compared using both current-control and voltage-control.
The current-control methods are initialized with a starting current of 2A and the voltage-control methods are initialized
with a starting voltage of 300V . Every method is able to track the 800W /m2 step-up, although the perturb and observe
methods have slower response. Only the voltage-control methods are able to track the step-down in irradiance at t = 40s.
At G = 1000W /m2 the current-control methods can only track a step-down of about 40W /m2 before short-circuit, as seen
in Figure 10. For this reason, both current-control methods fail to track the step-down.
V. CONCLUSIONS
A novel extremum seeking algorithm that utilizes the natural inverter ripple was tested on a simulated array-inverter
system. This method was benchmarked against the popular perturb and observe method using 25 minutes of rapidly varying
irradiance data taken on June, 2007 at Princeton University. The irradiance data represents a worst-case scenario for
maximum power point tracking due to the presence of fast moving, scattered cloud cover. It was shown that extremum
seeking slightly outperforms perturb and observe in total power efficiency, and drastically outperforms in transient risetime to the maximum power point, with two orders of magnitude speed-up. Moreover, extremum seeking has guaranteed
convergence and stability properties which are ideal for variable weather conditions and unmodeled dynamics.
The extremum seeking and perturb and observe algorithms are compared with voltage-control and current-control. The
relative performance of the two algorithms is similar for both voltage and current control implementations. However, it is
shown that the shape of the maximum power voltage vs. irradiance curve provides significant control benefits including
larger margins and an acceptable failure mode. In contrast, the maximum power current vs. irradiance curve is very close
to the short-circuit curve, leaving narrow failure margins which result in short-circuit. The voltage-control implementation
is fast enough to handle fully varying irradiance data, while the current-control implementation is only fast enough to track
irradiance data that is low-pass filtered over roughly 10s.
A major result of this work is that the ripple-based extremum seeking algorithm has good MPPT performance over a
range of inverter capacitor sizes. Typically the choice of capacitor is expensive because it must be well characterized and
large enough to maintain a small ripple. However, because the extremum seeking control signal exploits the natural inverter
ripple, a smaller capacitor allows the tracking of rapid irradiance changes. Additionally, the extremum seeking algorithm may
13
be built using analog components and wrapped around an existing array-inverter system with a voltage-control input. This
may influence inverter manufacturers to provide a voltage-control input.
ACKNOWLEDGMENT
The authors gratefully acknowledge the support for this work from the New Jersey Commission on Science and Technology (NJCST). The authors would like to thank Mark Holveck, Erik Limpaecher, Darren Hammell and Swarnab Banerjee
from Princeton Power Systems for gathering irradiance data, as well as for helpful discussions about inverter models. We
also thank the anonymous referees for their valuable comments and suggestions.
APPENDIX
This appendix contains details of the PV array-inverter system, including values and definitions used in the lighted-diode
equations (1)-(3). Additionally, information related to tuning the extremum seeking (ES) controller is included.
The system simulated consists of a PV array and grid-tied inverter. The array model consists of 3 parallel strings, each
with 7 panels connected in series. Each panel produces approximately 220W at full irradiance, G = 1000W /m2 . Unless
otherwise stated, the inverter hardware uses a 2000µF capacitor and 1µH inductor.
Values and definitions of the terms in equations (1)-(3) are as follows: TR = 298 (reference temperature), IOR = 2.25e − 6
(reverse saturation current at T = TR ), ISC = 3.2 (short-circuit current), EG = 1.8e − 19 (Silicon band gap), A = 1.6 (ideality
factor), kB = 1.38e − 23 (Boltzmann’s constant), q = 1.6e − 19 (electronic charge), R = .01 (resistance), and KT,I = .8
(short-circuit current temperature coefficient).
Tuning the extremum seeking controller typically involves choice of the perturbation size α and frequency ω as well
as the high-pass filter cut-off frequency ωh and integrator gain k. By utilizing the inverter ripple for the perturbation, α is
determined by the capacitor size and ω is given by twice the inverter AC frequency; 60Hz demodulated is 120Hz. The cutoff frequency ωh of the high-pass filter must be slower than the periodic perturbation, and the integrator gain k should be
small, as discussed in [13]. In practice, ωh = 750rad/sec works well, even though it is very close to 120Hz. The integrator
gain k is .2 for current-control and .3 for voltage-control, and it is chosen so that the response to transients is fast and stable.
In general, k must be larger if the capacitor is large, because the magnitude of the ripple-base control signal, is smaller
for a large capacitor. However, the extremum seeking algorithm is adaptive and does not require fine-tuning for excellent
performance. In contrast, the step size of the perturb and observe method must be carefully chosen to balance the tradeoffs
between rise-time and steady-state performance depending on the variability in irradiance.
REFERENCES
14
[1] V. Quaschning and R. Hanitsch, ”Influence of shading on electrical parameters of solar cells,” Photovoltaic Specialists Conference,
1996., Conference Record of the Twenty Fifth IEEE, pp. 1287–1290, May 1996.
[2] K. Hussein, I. Muta, T. Hoshino, and M. Osakada, ”Maximum photovoltaic power tracking - an algorithm for rapidly changing
atmospheric conditions,” Iee Proceedings-Generation Transmission and Distribution, vol. 142, pp. 59–64, January 1995.
[3] G. Vachtsevanos and K. Kalaitzakis, ”A Hybrid Photovoltaic Simulator for Utility Interactive Studies,” IEEE Transactions on Energy
Conversion, EC-2 (2), pp. 227-231, 1987.
[4] B. Bose, P. Szczesny, and R. Steigerwald, ”Microcomputer control of a residential photovoltaic power conditioning system,” IEEE
Transactions On Industry Applications, vol. 21, pp. 1182–1191, 1985.
[5] B.K. Bose, Modern Power Electronics: Evolution, Technology, and Applications. New York: IEEE Press, 1992.
[6] R. A. Mastromauro, M. Liserre, T. Kerekes, and A. Dell’Aquila, ”A single-phase voltage-controlled grid-connected photovoltaic system
with power quality conditioner functionality,” IEEE Transactions On Industrial Electronics, vol. 56, pp. 4436–4444, November 2009.
[7] Y. Porasad and H. Hosseinzadeh, ”Comparison of voltage control and current control methods in grid connected inverters,” Journal
of Applied Sciences, vol. 8, pp. 648–653, 2008.
[8] K. Tse, B. Ho, H. Chung, and S. Hui, ”A comparative study of maximum-power-point trackers for photovoltaic panels using switchingfrequency modulation scheme,” IEEE Transactions On Industrial Electronics, vol. 51, pp. 410–418, April 2004.
[9] D. Hohm and M. Ropp, ”Comparative study of maximum power point tracking algorithms,” Progress In Photovoltaics, vol. 11, pp.
47–62, January 2003.
[10] T. Esram and P. L. Chapman, ”Comparison of photovoltaic array maximum power point tracking techniques,” IEEE Transactions On
Energy Conversion, vol. 22, pp. 439–449, 2007.
[11] N. Femia, G. Petrone, G. Spagnuolo, and M. Vitelli, ”Optimization of perturb and observe maximum power point tracking method,”
IEEE Transactions On Power Electronics, vol. 20, pp. 963–973, 2005.
[12] N. Femia, D. Granozio, G. Petrone, G. Spagnuolo, and M. Vitelli, ”Predictive & adaptive mppt perturb and observe method,” IEEE
Transactions On Aerospace and Electronic Systems, vol. 43, pp. 934–950, 2007.
[13] M. Krstic and H. Wang, ”Stability of extremum seeking feedback for general nonlinear dynamic systems,” Automatica, vol. 36, pp.
595–601, April 2000.
[14] R. Leyva, C. Alonso, I. Queinnec, A. Cid-Pastor, D. Lagrange, and L. Martinez-Salamero, ”Mppt of photovoltaic systems using
extremum-seeking control,” IEEE Transactions On Aerospace and Electronic Systems, vol. 42, pp. 249–258, January 2006.
[15] A. I. Bratcu, I. Munteanu, S. Bacha, and B. Raison, ”Maximum power point tracking of grid-connected photovoltaic arrays by using
extremum seeking control,” CEAI, vol. 10, pp. 3–12, 2008.
[16] D. Tokushima, M. Uchida, S. Kanbei, H. Ishikawa, and H. Naitoh, ”A new mppt control for photovoltaic panels by instantaneous
maximum power point tracking,” Electrical Engineering In Japan, vol. 157, pp. 73–80, November 2006.
[17] P. Midya, P. Krein, and R. Turnbull, ”Dynamic maximum power point tracker for photovoltaic applications,” In Power Electronics
Specialists Conference, 1996. PESC ’96 Record., 27th Annual IEEE, vol. 2, June 1996.
[18] T. Esram, J. W. Kimball, P. T Krein, P. L. Chapman, and P. Midya, ”Dynamic maximum power point tracking of photovoltaic arrays
using ripple correlation control,” IEEE Transactions On Power Electronics, vol. 21, pp. 1282–1291, September 2006.
15
[19] J. W. Kimball, and P. T. Krein, ”Discrete-time ripple correlation control for maximum power point tracking,” IEEE Transactions on
Power Electronics, vol. 23, pp. 2353–2362, September 2008.
[20] J. Choi, M. Krstic, K. Ariyur, and J. Lee, ”Extremum seeking control for discrete-time systems,” IEEE Transactions On Automatic
Control, vol. 47, pp. 318–323, February 2002.
[21] Y. H. Lim and D. Hamill, ”Simple maximum power point tracker for photovoltaic arrays,” Electronics Letters, vol. 36, pp. 997–999,
May 2000.
[22] Y. H. Lim and D. Hamill, ”Synthesis, simulation and experimental verification of a maximum power point tracker from nonlinear
dynamics,” Power Electronics Specialists Conference, 2001. PESC. 2001 IEEE 32nd Annual, vol. 1, pp. 199–204, 2001.
16
Power (W)
6000
5000
MPP
ES
PO
4000
3000
2000
0
500
1000
1500
1000
1500
1000
1500
Current (I)
Time (s)
15
10
5
0
500
Time (s)
Voltage (V)
400
350
300
0
500
Time (s)
Fig. 6. Comparison of voltage-controlled extremum seeking (ES, black) and perturb and observe (PO, light gray) controllers on
25 minutes of unfiltered irradiance data from 12:34-12:59PM. (top) Both controllers track the MPP (dark gray). Commanded array
current (middle) and array voltage (bottom) show the large variation in maximum power current (MPI) with irradiance and the relative
stability of the maximum power voltage (MPV).
,
012+3,-4/
'!!!
"!!!
&!!!
500
67
08
%!!!
$!!! ,
!
"!!
#!!!
#"!!
#!!!
#"!!
#!!!
#"!!
9:33+;<,-=/
()*+,-./
#"
#!
"
!
"!!
()*+,-./
>1?<@A+,->/
&!!
%"!
%!!
!
"!!
()*+,-./
Fig. 7. Comparison of current-controlled extremum seeking (ES, black) and perturb and observe (PO, light gray) controllers on
25 minutes of low-pass filtered irradiance data from 12:34-12:59PM. (top) Both controllers track the MPP (dark gray). Commanded
array current (middle) and array voltage (bottom) show the large variation in maximum power current (MPI) and the relative stability
of the maximum power voltage (MPV).
17
6000
Power (W)
5000
4000
3000
2000
1000
0
0
20
MPP
PO 6I = 1.0
500
PO 6I
= .50
PO 6I = .25
PO 6I = .10
PO 6I = .05
1000
1500
1000
1500
Current (A)
15
10
5
0
0
500
Time (s)
Fig. 8. Comparison of perturb and observe (PO) controllers, using current-control, with varying perturbation step size ∆I. As ∆I is
increased, the rise-time decreases and oscillations about the MPP increase. The only PO algorithm which is able to track the MPP
for every irradiance change is ∆I = .5A.
18
6000
Power (W)
5000
4000
MPP
Cap = 500µF
Cap = 1000µF
Cap = 2000µF
Cap = 3000µF
3000
2000
1000
0
0
500
1000
1500
1000
1500
Time (s)
20
Current (A)
15
10
5
0
0
500
Time (s)
Fig. 9. Comparison of extremum seeking (ES) controllers, using current-control, on inverters with varying capacitor sizes. As
capacitor size increases the ES algorithm tracks the MPP with less oscillation, due to the decrease in ripple magnitude. However, if
the capacitor is too large the ripple-based control signal is insufficient to track rapid irradiance changes, and the algorithm fails, as
seen when the capacitor is 3000µF.
Maximum Power Voltage, Open Circuit Voltage
1000
Maximum Power Current, Short Circuit Current
1000
MPV
Open Circuit
800
Irradiance, G
Irradiance, G
800
600
400
200
0
250
MPI
Short Circuit
600
400
200
300
350
Voltage, V
400
450
0
0
5
10
15
Current, I
Fig. 10. (left) Maximum power voltage (solid) and open-circuit voltage plotted against irradiance G. (right) Maximum power current
(solid) and short-circuit current plotted against irradiance G.
19
'$'%
'$')%
2-,345678-4.=5::->8<.?@A
!"#$%&'(()*+,-.-/
2-,345678-4.9:345;8<.!
'
!'$'%
!'$)
!'$)%
!'$"
!'$"%
!'$(
!'$(%
!"#
!"#$%
!"&
*+,-./01
!"&$%
!"!
'$')
'$''%
'
!'$''%
!'$')
!'$')%
!"#
!"#$%
!"&
*+,-./01
!"&$%
!"!
Fig. 11. (left) Demodulated signal, ξ = Ihp × Php . (right) High-pass filtered current, Ihp . As irradiance falls, average current Iˆ nears
short-circuit, the ripple is constrained, and the amplitude decreases as the method fails.
Fig. 12. (left) Power vs. time for maximum power (gray) and extremum seeking (black). (right) Corresponding P-I trace at four
stages of tracking: (1) rise, (2) increasing irradiance, (3) decreasing irradiance, and (4) failure.
()*+,-./0
20
%!!!
#!!!
!
!
"!
#!
$!
%!
&!
"&
12,,+34-.50
'!
"!
&
!
!
"!
#!
:);4<=+-.:0
%&!
$!
%!
'!
&!
'!
!"#$%&'
%!!
&!
$&!
$!!
#&!
!
"!
#!
$!
678+-.90
%!
Fig. 13. Comparison of extremum seeking (ES) and perturb and observe (PO) using current-control (-I) and voltage-control (-V)
on step irradiance changes. The irradiance steps up at t = 20s from 200W /m2 to 1000W /m2 and steps back down to 200W /m2 at
t = 40s. Only voltage-control methods track the step-down.